Kimi K2
1T-parameter (32B active) open-weights MoE trained with the MuonClip optimizer on 15.5T tokens, with zero reported loss spikes and a focus on agentic tool use.
First open-weights model past 1T parameters, aimed squarely at agents. It uses MuonClip (Muon plus QK-clip) for stable training, large-scale synthetic agentic tool-use data, and joint RL with real and simulated environments. Non-thinking model.
- Date
- Friday, 11 July 2025
- Lab
- Moonshot AI
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-bench Verified (agentic, single attempt) | 65.8% bash/editor tools, no test-time compute | company |
| Tau2-Bench | 66.1 | company |
| LiveCodeBench v6 | 53.7 | company |
| GPQA-Diamond | 75.1 | company |
Modified MIT license, with attribution required above 100M MAU or $20M monthly revenue. 128K context. Paper on arXiv 2025-07-28. Release date from HF repo commits (2025-07-11) and Epoch AI.
Sources
- huggingface.co/moonshotai/Kimi-K2-Instruct
- arxiv.org/abs/2507.20534
- simonwillison.net/tags/ai-in-china/?page=2
This record was checked against its sources on 6 October 2026. How we check